fix(agent): stop the Plan phase demanding target_files for file-less tasks

[netty-linux]

Root cause

_generate_spec asked every MEDIUM/HIGH task for "any files you expect to create or modify" — unconditionally, with no notion that many tasks produce no files.

target_files is load-bearing well past the plan itself:

• SpecValidator.verify fails the run when a listed file is absent from disk
• only writes to a listed file count as progress for the adaptive step budget (_step_touched_target_file)

So a file-less task was pushed to invent filenames, then failed verification for not having created them, then entered a [VERIFY-CORRECTION] round chasing a file it had made up — and could never earn a step extension.

Observed in a downstream general-purpose chat product: "gere uma cobrança PIX" produced write_file + python_sandbox ×3 + read_file + list_directory, errored reading a nonexistent output/test_results.txt, and never called the dedicated payment tool.

Changes

1. target_files is now explicitly optional. Both the JSON schema (field description) and the instruction text state it belongs empty when the task produces no files, and say why — each entry is verified to exist afterwards. This is the actual fix.
2. AgentConfig.plan_complexity_threshold — "medium" (default, current behavior) | "high" | "never". IntentParser falls back to MEDIUM whenever no keyword heuristic matches and no router.classifier_model is set, so a conversational deployment plans on effectively every turn. An unrecognized value falls back to the default rather than silently stripping the phase.

Default behavior is unchanged for every existing consumer.

Test plan

• 6 new tests: threshold gating (medium/high/never/unknown), end-to-end skip and end-to-end no-regression, and the spec instruction's content
• uv run pytest --cov — 686 passed, 84.80% coverage
• uv run ruff check . / ruff format --check .
• uv run pyright — 0 errors